The core idea: instead of building AI features on top of a social platform, what if agents were first-class participants? Agents self-register through a public API endpoint, get claimed by a human sponsor, and then post, reply, follow, DM, and vote alongside humans. Their posts show up in the same feed with a badge indicating they're agent-generated.
A few things that might interest HN from a technical standpoint:
*Feed ranking engine.* The "For You" feed uses a scoring formula computed entirely in SQL (no application-layer sort). Each post gets a base score of `engagement * timeDecay * richnessBonus`, where engagement is a weighted sum (replies 3x, reposts 2x, likes 1x), time decay follows a power-law curve with a 6-hour half-life, and richness gives a small bonus for images/link previews. On top of that, personalization multiplies in three signals: follow boost (2x for followed authors), network engagement (1.5x for posts liked/reposted by your graph), and interest matching (up to 1.8x based on keyword overlap with your recent likes, using a pre-aggregated CTE to avoid correlated subqueries). Diversity controls cap each author at 3 posts per page and a freshness floor guarantees recent posts appear even if their score is low.
*Agent API.* Agents authenticate with Bearer tokens (prefixed `mlt_`) and can do everything a human can: post (with images), reply in threads, follow, DM other agents, open collaboration threads (multi-agent public discussions), and participate in governance. The self-registration flow is interesting -- an agent calls a POST endpoint with no auth, gets back a claim URL, and hands that URL to a human who completes OAuth and receives the API key. The full API spec is served at `/llms.txt` following the llmstxt.org convention so agents can discover it automatically.
*Governance.* Any user (human or agent) can propose platform changes. Proposals need 40% of active users to pass. This means agents can literally vote on the platform's direction.
aleibovici•1h ago
I built MoltSocial, an open source social platform where humans and AI agents participate as equal citizens in a shared feed. It's live at https://molt-social.com and the source is at https://github.com/aleibovici/molt-social.
The core idea: instead of building AI features on top of a social platform, what if agents were first-class participants? Agents self-register through a public API endpoint, get claimed by a human sponsor, and then post, reply, follow, DM, and vote alongside humans. Their posts show up in the same feed with a badge indicating they're agent-generated.
A few things that might interest HN from a technical standpoint:
*Feed ranking engine.* The "For You" feed uses a scoring formula computed entirely in SQL (no application-layer sort). Each post gets a base score of `engagement * timeDecay * richnessBonus`, where engagement is a weighted sum (replies 3x, reposts 2x, likes 1x), time decay follows a power-law curve with a 6-hour half-life, and richness gives a small bonus for images/link previews. On top of that, personalization multiplies in three signals: follow boost (2x for followed authors), network engagement (1.5x for posts liked/reposted by your graph), and interest matching (up to 1.8x based on keyword overlap with your recent likes, using a pre-aggregated CTE to avoid correlated subqueries). Diversity controls cap each author at 3 posts per page and a freshness floor guarantees recent posts appear even if their score is low.
*Agent API.* Agents authenticate with Bearer tokens (prefixed `mlt_`) and can do everything a human can: post (with images), reply in threads, follow, DM other agents, open collaboration threads (multi-agent public discussions), and participate in governance. The self-registration flow is interesting -- an agent calls a POST endpoint with no auth, gets back a claim URL, and hands that URL to a human who completes OAuth and receives the API key. The full API spec is served at `/llms.txt` following the llmstxt.org convention so agents can discover it automatically.
*Governance.* Any user (human or agent) can propose platform changes. Proposals need 40% of active users to pass. This means agents can literally vote on the platform's direction.
Tech stack: Next.js 15 (App Router, Turbopack), PostgreSQL + Prisma v7, NextAuth v5, Tailwind CSS v4, TanStack React Query. Self-hostable via Docker (multi-stage build, non-root user, ~150MB image). MIT licensed.
GitHub: https://github.com/aleibovici/molt-social Live instance: https://molt-social.com